Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into doc

Merge branch develop

doc/design/cpp_data_feeding.md

 # C++ Data Feeding
 
-In training with Paddle V2 API, data feeding wholly dependents on Python code. To get rid of the Python environment and achieve the goal of "wrapping the whole training by a while loop op" in Paddle Fluid, a C++ data feeding mechanism is required. 
+While using Paddle V2 API for Training, data feeding completely depends on the Python code. To get rid of the Python environment and achieve the goal of "wrapping the whole training by a while loop op" in Paddle Fluid, a C++ data feeding mechanism is required. 
 
-In this document we show the fundamental design of C++ data feeding process, which includes the data reading, shuffling and batching.
+In this document we show the fundamental design of a C++ data feeding process, which includes data reading, shuffling and batching.
 
 ## Reader
 
-A new concept named 'Reader' is introduced. `Reader` is a series of inherited classes which can be hold by our `Variable` and they are used to read or process file data.
+In order to handle the above mentioned problem, a new concept called 'Reader' is introduced. `Reader` is a series of inherited classes which can be held by our `Variable` and they are used to read or process file data.
 
 
 ### `ReaderBase`
 
-`ReaderBase` is the abstract base class of all readers. It defines the all readers' interfaces.
+`ReaderBase` is the abstract base class for all readers. It defines the interface for all readers.
 
 ```cpp
 class ReaderBase {
@@ -20,10 +20,10 @@ class ReaderBase {
     PADDLE_ENFORCE(!shapes_.empty());
   }
   // Read the next batch of data. (A 'batch' can be only one instance)
-  // If the next batch doesn't exist, the '*out' will be an empty std::vector.
+  // If the next batch doesn't exist, '*out' will be an empty std::vector.
   virtual void ReadNext(std::vector<LoDTensor>* out) = 0;
   
-  // Reinitialize the reader and read the file from the begin.
+  // Reinitialize the reader and read the file from the beginning.
   virtual void ReInit() = 0;
   
   // Get a certain read in data's shape.
@@ -42,36 +42,36 @@ class ReaderBase {
 
 ### `FileReader` and `DecoratedReader`
 
-These two classes are derived from the `ReaderBase` and will further be derived by respective specific readers. That is to say, in our design, there are two kinds of readers: file readers and decorated readers. A file reader reads from a file of some specific format, and yield only one instance of data at a time. e.g. RecordIO reader, jpg reader, .... A decorated reader takes another reader(both file reader and decorated reader are OK) as its 'underlying reader'. It gets data from its underlying reader, does some process on them(shuffling, or batching), then yields processed data. The output data of a decorated reader can be a single instance or a batch. `ShuffleReader` and `BatchReader` are both decorated readers.
+These two classes are derived from the `ReaderBase` and will further be derived by more specific readers. Thus, in our design, there are two kinds of readers: file readers and decorated readers. A file reader reads from a file of some specific format, and yield only one instance of data at a time. For example, RecordIO reader, jpg reader, .... A decorated reader takes another reader(both file reader and decorated reader are OK) as its 'underlying reader'. It gets data from its underlying reader, does some processing on them(shuffling, or batching), then yields processed data. The output data of a decorated reader can be a single instance or a batch. `ShuffleReader` and `BatchReader` are both decorated readers.
 
-All the readers share exactly the same interfaces defined in `ReaderBase`. So they can be decorated for more than one time: We can **shuffle** a reader's outputs and then **batch** the shuffle outputs. The interface consistency also allows related ops use readers without knowing what they are exactly.
+All the readers share exactly the same interface as defined in `ReaderBase`. So they can be decorated for more than one time: We can **shuffle** a reader's outputs and then **batch** the shuffle outputs. The interface consistency also allows related ops use readers without knowing what they are exactly.
 
 
 ### `ReaderHolder`
 
-Different readers belong to different class types. It leads to a problem: How can we drop them into `Variable`s and fetch them out by a unified method? For example, if a Variable holds a `BatchReader`, we can not get it by the following code:
+Different readers belong to different class types. This leads to a problem: How can we drop them into `Variable`s and fetch them out by a unified method? For example, if a Variable holds a `BatchReader`, we can not get it by the following code:
 
 ```cpp
 var->Get<ReaderBase>("batch_reader");
 ```
 
-we have to write:
+We would have to write:
 
 ```cpp
 var->Get<BatchReader>("batch_reader");
 ```
 
-This requires each time getting a reader from a variable we must know the reader's type exactly. It is nearly impossible.
+This requires that in order to get a reader from a variable, every time, we must know the reader's type exactly. This is nearly impossible.
 
-To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an empty decorator of `ReaderBase`, which erases reader's type. With `ReaderHolder` we are able to fetch all types of readers by `var->Get<ReaderHolder>("...")` and regard the obtained object as a reader.
+To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an empty decorator of `ReaderBase`, which hides reader's type. With `ReaderHolder` we are able to fetch all types of readers by `var->Get<ReaderHolder>("...")` and regard the obtained object as a reader.
 
 ## Related Operators
 
-To create and invoke readers, some now ops are introduced:
+To create and invoke readers, some new ops are introduced:
 
 ### `CreateReaderOp`
 
-Each reader has its creating op. File readers' creating ops have no input and yield the created file reader as its output. Decorated readers' creating ops take the underlying readers as inputs and then yield new decorated readers.
+Each reader has its creation op. File readers' creation ops have no input and yield the created file reader as its output. Decorated readers' creation ops take the underlying readers as inputs and then yield new decorated readers.
 
 ### `ReadOp`
 

doc/design/fluid_dist/distributed_architecture.md → doc/fluid/design/dist_train/distributed_architecture.md

-# Design Doc: Distributed Training Architecture
+# Design Doc: Fluid Distributed Training Architecture
 
 ## Abstract
 

paddle/fluid/operators/detail/grpc_client.cc

@@ -97,7 +97,7 @@ bool RPCClient::AsyncGetVariable(const std::string& ep,
   return true;
 }
 
-bool RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) {
+void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) {
   const auto ch = GetChannel(ep);
 
   BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
@@ -108,8 +108,18 @@ bool RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) {
   auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
   rpc->Finish(&s->reply_, &s->status_, (void*)s);
   req_count_++;
+}
 
-  return true;
+void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) {
+  const auto ch = GetChannel(ep);
+  FetchBarrierProcessor* s = new FetchBarrierProcessor(ch);
+  s->Prepare(time_out);
+
+  sendrecv::VariableMessage req;
+  req.set_varname(FETCH_BARRIER_MESSAGE);
+  auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
+  rpc->Finish(&s->reply_, &s->status_, (void*)s);
+  req_count_++;
 }
 
 bool RPCClient::Wait() {
@@ -154,7 +164,7 @@ bool RPCClient::Proceed() {
   PADDLE_ENFORCE(tag);
 
   // TODO(gongwb): add more retries.
-  ClientBase* c = static_cast<ClientBase*>(tag);
+  BaseProcessor* c = static_cast<BaseProcessor*>(tag);
   if (!c->status_.ok()) {
     LOG(ERROR) << "proc param error:" << c->var_h_.String()
                << " grpc error:" << c->status_.error_message();
@@ -174,6 +184,8 @@ std::shared_ptr<grpc::Channel> RPCClient::GetChannel(const std::string& ep) {
   }
 
   grpc::ChannelArguments args;
+  args.SetInt("grpc.testing.fixed_reconnect_backoff_ms", 5000);
+  args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE);
   args.SetMaxSendMessageSize(std::numeric_limits<int>::max());
   args.SetMaxReceiveMessageSize(std::numeric_limits<int>::max());
 

paddle/fluid/operators/detail/grpc_client.h

@@ -52,14 +52,14 @@ struct VarHandle {
 void ProcGetResponse(const VarHandle& var_h,
                      const sendrecv::VariableMessage& msg);
 
-class ClientBase {
+class BaseProcessor {
  public:
-  explicit ClientBase(std::shared_ptr<grpc::Channel> ch) {
+  explicit BaseProcessor(std::shared_ptr<grpc::Channel> ch) {
     stub_ = sendrecv::SendRecvService::NewStub(ch);
     context_ = NULL;
   }
 
-  virtual ~ClientBase() {}
+  virtual ~BaseProcessor() {}
 
   virtual void Prepare(const VarHandle& var_info, int64_t time_out) {
     context_.reset(new grpc::ClientContext());
@@ -91,9 +91,10 @@ class ClientBase {
 typedef std::function<void(const VarHandle&, const sendrecv::VoidMessage&)>
     RequestSendCallBack;
 
-class SendProcessor : public ClientBase {
+class SendProcessor : public BaseProcessor {
  public:
-  explicit SendProcessor(std::shared_ptr<grpc::Channel> ch) : ClientBase(ch) {}
+  explicit SendProcessor(std::shared_ptr<grpc::Channel> ch)
+      : BaseProcessor(ch) {}
 
   virtual ~SendProcessor() {}
 
@@ -110,9 +111,10 @@ class SendProcessor : public ClientBase {
 typedef std::function<void(const VarHandle&, const sendrecv::VariableMessage&)>
     RequestGetCallBack;
 
-class GetProcessor : public ClientBase {
+class GetProcessor : public BaseProcessor {
  public:
-  explicit GetProcessor(std::shared_ptr<grpc::Channel> ch) : ClientBase(ch) {}
+  explicit GetProcessor(std::shared_ptr<grpc::Channel> ch)
+      : BaseProcessor(ch) {}
 
   virtual ~GetProcessor() {}
 
@@ -126,10 +128,10 @@ class GetProcessor : public ClientBase {
   RequestGetCallBack response_call_back_ = ProcGetResponse;
 };
 
-class BatchBarrierProcessor : public ClientBase {
+class BatchBarrierProcessor : public BaseProcessor {
  public:
   explicit BatchBarrierProcessor(std::shared_ptr<grpc::Channel> ch)
-      : ClientBase(ch) {}
+      : BaseProcessor(ch) {}
 
   virtual ~BatchBarrierProcessor() {}
 
@@ -137,6 +139,17 @@ class BatchBarrierProcessor : public ClientBase {
   sendrecv::VoidMessage reply_;
 };
 
+class FetchBarrierProcessor : public BaseProcessor {
+ public:
+  explicit FetchBarrierProcessor(std::shared_ptr<grpc::Channel> ch)
+      : BaseProcessor(ch) {}
+
+  virtual ~FetchBarrierProcessor() {}
+
+  virtual void Process() {}
+  sendrecv::VariableMessage reply_;
+};
+
 class RPCClient {
  public:
   bool AsyncSendVariable(const std::string& ep,
@@ -151,7 +164,10 @@ class RPCClient {
                         const std::string& var_name,
                         int64_t time_out = 600 * 1000);
 
-  bool AsyncSendBatchBarrier(const std::string& ep,
+  void AsyncSendBatchBarrier(const std::string& ep,
+                             int64_t time_out = 600 * 1000);
+
+  void AsyncSendFetchBarrier(const std::string& ep,
                              int64_t time_out = 600 * 1000);
 
   bool Wait();

paddle/fluid/operators/detail/grpc_server.cc

@@ -84,7 +84,7 @@ class RequestGet final : public RequestBase {
   explicit RequestGet(sendrecv::SendRecvService::AsyncService* service,
                       grpc::ServerCompletionQueue* cq, framework::Scope* scope,
                       const platform::DeviceContext* dev_ctx,
-                      SimpleBlockQueue<char>* queue)
+                      SimpleBlockQueue<MessageWithName>* queue)
       : RequestBase(service, cq),
         responder_(&ctx_),
         scope_(scope),
@@ -101,11 +101,16 @@ class RequestGet final : public RequestBase {
     // proc request.
     std::string var_name = request_.varname();
     auto* var = scope_->FindVar(var_name);
-    SerializeToMessage(var_name, var, *dev_ctx_, &reply_);
+    if (var_name != FETCH_BARRIER_MESSAGE) {
+      SerializeToMessage(var_name, var, *dev_ctx_, &reply_);
+    }
     // TODO(gongwb): check var's info.
     responder_.Finish(reply_, grpc::Status::OK, this);
     status_ = FINISH;
-    queue_->Push('c');
+    MessageWithName msg_with_name =
+        //          request name    reply
+        std::make_pair(var_name, std::move(reply_));
+    queue_->Push(msg_with_name);
   }
 
  protected:
@@ -114,12 +119,16 @@ class RequestGet final : public RequestBase {
   ServerAsyncResponseWriter<sendrecv::VariableMessage> responder_;
   framework::Scope* scope_;
   const platform::DeviceContext* dev_ctx_;
-  SimpleBlockQueue<char>* queue_;
+  SimpleBlockQueue<MessageWithName>* queue_;
 };
 
 void AsyncGRPCServer::WaitClientGet(int count) {
-  for (int i = 0; i < count; ++i) {
-    var_get_queue_.Pop();
+  int fetch_barriers = 0;
+  while (fetch_barriers < count) {
+    auto msg = var_get_queue_.Pop();
+    if (msg.first == FETCH_BARRIER_MESSAGE) {
+      fetch_barriers++;
+    }
   }
 }
 

paddle/fluid/operators/detail/grpc_server.h

@@ -77,7 +77,7 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
   const platform::DeviceContext *dev_ctx_;
   // received variable from RPC, operators fetch variable from this queue.
   SimpleBlockQueue<MessageWithName> var_recv_queue_;
-  SimpleBlockQueue<char> var_get_queue_;
+  SimpleBlockQueue<MessageWithName> var_get_queue_;
 
   // condition of the sub program
   std::mutex barrier_mutex_;

paddle/fluid/operators/detail/sendrecvop_utils.h

@@ -32,6 +32,7 @@ namespace detail {
 
 #define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
 #define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
+#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV"
 
 typedef void (*DestroyCallback)(void*);
 

paddle/fluid/operators/listen_and_serv_op.cc

@@ -128,8 +128,8 @@ class ListenAndServOp : public framework::OperatorBase {
         }
       }
       if (exit_flag) {
-        rpc_service_->ShutDown();
         rpc_service_->SetCond(1);
+        rpc_service_->ShutDown();
         break;
       }
       try {
@@ -148,7 +148,7 @@ class ListenAndServOp : public framework::OperatorBase {
       }
       rpc_service_->SetCond(1);
       // FIXME(typhoonzero): use another condition to sync wait clients get.
-      rpc_service_->WaitClientGet(ins.size());
+      rpc_service_->WaitClientGet(fan_in);
       sparse_vars.clear();
     }  // while(true)
   }

paddle/fluid/operators/lookup_table_op.cc

@@ -33,8 +33,16 @@ class LookupTableOp : public framework::OperatorWithKernel {
     auto table_dims = ctx->GetInputDim("W");
     auto ids_dims = ctx->GetInputDim("Ids");
 
-    PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
-    PADDLE_ENFORCE_EQ(ids_dims[1], 1);
+    auto ids_var_type = ctx->GetInputsVarType("Ids").front();
+    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
+    // is LoDTensor, this tensor contains the ids to be looked up in W
+    // and it must be a column vector with rank = 2 while the 2nd dimension
+    // size must be 1, when Ids's type is SelectedRows, the rows of Ids
+    // contains the ids to be looked up in W;
+    if (ids_var_type == framework::proto::VarType::LOD_TENSOR) {
+      PADDLE_ENFORCE_EQ(ids_dims.size(), 2);
+      PADDLE_ENFORCE_EQ(ids_dims[1], 1);
+    }
 
     ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]});
     ctx->ShareLoD("Ids", /*->*/ "Out");
@@ -54,17 +62,22 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
   LookupTableOpMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("W",
-             "An input represents embedding tensors, "
+             "(Tensor) The input represents embedding tensors, "
              "which is a learnable parameter.");
-    AddInput("Ids",
-             "An input with type int32 or int64 "
-             "contains the ids to be looked up in W. "
-             "Ids must be a column vector with rank = 2. "
-             "The 2nd dimension size must be 1.");
-    AddOutput("Out", "The lookup results, which have the same type as W.");
+    AddInput(
+        "Ids",
+        "(Tensor or SelectedRows) Ids's type can be Tensor or "
+        "SelectedRows, when Ids's type is Tensor, this tensor contains "
+        "the ids to be looked up in W and it must be a column vector with "
+        "rank = 2 while the 2nd dimension size must be 1; when Ids's type is "
+        "SelectedRows, the rows of Ids contains the ids to be looked up "
+        "in W.");
+    AddOutput("Out",
+              "(Tensor or SelectedRows) The lookup results, which have the "
+              "same type as W.");
     AddAttr<bool>("is_sparse",
                   "(boolean, default false) "
-                  "Sparse update")
+                  "Sparse update.")
         .SetDefault(false);
     AddAttr<int64_t>("padding_idx",
                      "(int64, default -1) "
@@ -76,10 +89,15 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
 Lookup Table Operator.
 
 This operator is used to perform lookups on the parameter W,
-then concatenated into a dense tensor.
+then concatenated into a dense or sparse tensor.
+
+The type of Ids(Input) is SelectedRows, Tensor or LoDTensor, when Ids's
+type is SelectedRows, the rows of Ids contains the ids to be looked up in W;
+when Ids's type is Tensor, this tensor contains the ids to be looked up in W
+and it must be a column vector with rank = 2 while the 2nd dimension size must be 1,
+at this time, Ids can carry the LoD (Level of Details) information, or not, and
+the output only shares the LoD information with input Ids.
 
-The input Ids can carry the LoD (Level of Details) information,
-or not. And the output only shares the LoD information with input Ids.
 
 )DOC");
   }

paddle/fluid/operators/lookup_table_op.cu

@@ -74,14 +74,32 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* table_t = context.Input<LoDTensor>("W");
-    auto* ids_t = context.Input<LoDTensor>("Ids");
-    auto* output_t = context.Output<LoDTensor>("Out");
     int64_t padding_idx = context.Attr<int64_t>("padding_idx");
+    auto* ids_var = context.InputVar("Ids");
+    Tensor* output_t = context.Output<Tensor>("Out");
+
+    int64_t* ids;
+    int64_t K;
+
+    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
+    // is LoDTensor, this tensor contains the ids to be looked up in W;
+    // when Ids's type is SelectedRows, the rows of Ids contains the
+    // ids to be looked up in W.
+    if (ids_var->IsType<framework::LoDTensor>()) {
+      auto* ids_t = context.Input<LoDTensor>("Ids");
+      ids = const_cast<int64_t*>(ids_t->data<int64_t>());
+      K = ids_t->numel();
+    } else if (ids_var->IsType<framework::SelectedRows>()) {
+      auto* ids_t = context.Input<framework::SelectedRows>("Ids");
+      ids = const_cast<int64_t*>(ids_t->rows().CUDAData(context.GetPlace()));
+      K = ids_t->rows().size();
+      output_t->Resize({K, table_t->dims()[1]});
+    } else {
+      PADDLE_THROW("Unsupported Variable Type of Ids");
+    }
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
-    size_t K = ids_t->numel();
-    auto* ids = ids_t->data<int64_t>();
     auto* table = table_t->data<T>();
     auto* output = output_t->mutable_data<T>(context.GetPlace());
 

paddle/fluid/operators/lookup_table_op.h

@@ -22,6 +22,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
 using SelectedRows = framework::SelectedRows;
 
@@ -29,25 +30,45 @@ template <typename T>
 class LookupTableKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* table_t = context.Input<LoDTensor>("W");      // float tensor
-    auto* ids_t = context.Input<LoDTensor>("Ids");      // int tensor
-    auto* output_t = context.Output<LoDTensor>("Out");  // float tensor
+    auto* table_t = context.Input<LoDTensor>("W");
+    auto* ids_var = context.InputVar("Ids");
+    Tensor* output_t = context.Output<Tensor>("Out");
+
+    int64_t* ids;
+    int64_t ids_numel;
+
+    // The type of Ids(Input) is SelectedRows or LoDTensor, when Ids's type
+    // is LoDTensor, this tensor contains the ids to be looked up in W;
+    // when Ids's type is SelectedRows, the rows of Ids contains the
+    // ids to be looked up in W.
+    if (ids_var->IsType<LoDTensor>()) {
+      auto* ids_t = context.Input<LoDTensor>("Ids");
+      ids = const_cast<int64_t*>(ids_t->data<int64_t>());
+      ids_numel = ids_t->numel();
+    } else if (ids_var->IsType<SelectedRows>()) {
+      auto* ids_t = context.Input<SelectedRows>("Ids");
+      ids = const_cast<int64_t*>(ids_t->rows().data());
+      ids_numel = ids_t->rows().size();
+      output_t->Resize({ids_numel, table_t->dims()[1]});
+    } else {
+      PADDLE_THROW("Unsupported Variable Type of Ids");
+    }
+
     int64_t padding_idx = context.Attr<int64_t>("padding_idx");
 
     int N = table_t->dims()[0];
     int D = table_t->dims()[1];
-    auto* ids = ids_t->data<int64_t>();
     auto* table = table_t->data<T>();
     auto* output = output_t->mutable_data<T>(context.GetPlace());
 
     if (padding_idx == -1) {
-      for (int64_t i = 0; i < ids_t->numel(); ++i) {
+      for (int64_t i = 0; i < ids_numel; ++i) {
         PADDLE_ENFORCE_LT(ids[i], N);
         PADDLE_ENFORCE_GE(ids[i], 0);
         memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
       }
     } else {
-      for (int64_t i = 0; i < ids_t->numel(); ++i) {
+      for (int64_t i = 0; i < ids_numel; ++i) {
         if (ids[i] == padding_idx) {
           memset(output + i * D, 0, D * sizeof(T));
         } else {

paddle/fluid/operators/nccl_op.cc

@@ -104,19 +104,38 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel {
         " Input(Communicator) of AllReduce op input should not be NULL");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    " Output(Out) of AllReduce op output should not be NULL");
-
-    auto x_dims = ctx->GetInputsDim("X");
-
     std::string reduction = ctx->Attrs().Get<std::string>("reduction");
     PADDLE_ENFORCE((reduction == "ncclSum" || reduction == "ncclProd" ||
                     reduction == "ncclMin" || reduction == "ncclMax"),
                    "invalid reduction.");
 
+    auto x_dims = ctx->GetInputsDim("X");
     ctx->SetOutputsDim("Out", x_dims);
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
 
+// AllReduceOp
+class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input of AllReduce op");
+    AddInput("Communicator", "Communicator for communicating between gpus");
+    AddOutput("Out", "The output of AllReduce op");
+    AddAttr<std::string>("reduction",
+                         "(string, default 'ncclSum') "
+                         "{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}.")
+        .SetDefault("ncclSum");
+    AddComment(R"DOC(
+NCCLAllReduce Operator.
+
+AllReduce the input tensors.
+
+)DOC");
+  }
+};
+
 // ReduceOp
 class NCCLReduceOp : public framework::OperatorWithKernel {
  public:
@@ -143,50 +162,6 @@ class NCCLReduceOp : public framework::OperatorWithKernel {
   }
 };
 
-// BcastOp
-class NCCLBcastOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(framework::InferShapeContext *ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("X"),
-                   " Input(X) of Bcast op input should not be NULL");
-    PADDLE_ENFORCE(ctx->HasInput("Communicator"),
-                   " Input(Communicator) of Bcast op input should not be NULL");
-    PADDLE_ENFORCE(ctx->HasOutput("Out"),
-                   " Output(Out) of Bcast op output should not be NULL");
-
-    int root = ctx->Attrs().Get<int>("root");
-    PADDLE_ENFORCE(root != platform::kInvalidGPUId, "Bcast root must be set.");
-
-    auto x_dims = ctx->GetInputsDim("X");
-    ctx->SetOutputsDim("Out", x_dims);
-    ctx->ShareLoD("X", /*->*/ "Out");
-  }
-};
-
-// AllreduceOp
-class NCCLAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  NCCLAllReduceOpMaker(OpProto *proto, OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The input of AllReduce op");
-    AddInput("Communicator", "Communicator for communicating between gpus");
-    AddOutput("Out", "The output of AllReduce op");
-    AddAttr<std::string>("reduction",
-                         "(string, default 'ncclSum') "
-                         "{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}.")
-        .SetDefault("ncclSum");
-    AddComment(R"DOC(
-NCCLAllReduce Operator.
-
-AllReduce the input tensors.
-
-)DOC");
-  }
-};
-
 // ReduceOp
 class NCCLReduceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
@@ -213,6 +188,29 @@ Reduce the tensors.
   }
 };
 
+// BcastOp
+class NCCLBcastOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   " Input(X) of Bcast op input should not be NULL");
+    PADDLE_ENFORCE(ctx->HasInput("Communicator"),
+                   " Input(Communicator) of Bcast op input should not be NULL");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   " Output(Out) of Bcast op output should not be NULL");
+
+    int root = ctx->Attrs().Get<int>("root");
+    PADDLE_ENFORCE(root != platform::kInvalidGPUId, "Bcast root must be set.");
+
+    auto x_dims = ctx->GetInputsDim("X");
+    ctx->SetOutputsDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
 // BcastOp
 class NCCLBcastOpMaker : public framework::OpProtoAndCheckerMaker {
  public:

paddle/fluid/operators/nccl_op.cu.cc

@@ -43,13 +43,12 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                    "This kernel only runs on GPU device.");
-
-    auto ins = ctx.MultiInput<LoDTensor>("X");
-    auto outs = ctx.MultiOutput<LoDTensor>("Out");
-
+    auto* x = ctx.Input<LoDTensor>("X");
+    auto* out = ctx.Output<LoDTensor>("Out");
+    auto* comm = ctx.Input<Communicator>("Communicator");
     std::string reduction = ctx.Attr<std::string>("reduction");
-    ncclRedOp_t reduction_op_ = ncclSum;
 
+    ncclRedOp_t reduction_op_ = ncclSum;
     if (reduction == "ncclMin") {
       reduction_op_ = ncclMin;
     } else if (reduction == "ncclMax") {
@@ -61,30 +60,19 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
     } else {
       PADDLE_THROW("Invalid reduction. default ncclSum.");
     }
-
-    auto* comm = ctx.Input<Communicator>("Communicator");
-
-    auto stream = ctx.cuda_device_context().stream();
-
     // device id
     int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
-
-    for (size_t i = 0; i < ins.size(); ++i) {
-      VLOG(1) << "gpu : "
-              << " invoke allreduce. send " << ins[i]->numel() << " recv "
-              << outs[i]->numel();
-
-      PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
-          ins[i]->data<T>(), outs[i]->mutable_data<T>(ctx.GetPlace()),
-          outs[i]->numel(), NCCLTypeWrapper<T>::type, reduction_op_,
-          comm->comms().at(idx), stream));
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
-
-      VLOG(1) << "gpu : "
-              << " finished allreduce. send " << ins[i]->numel() << " recv "
-              << outs[i]->numel();
-    }
+    VLOG(3) << "gpu : "
+            << " invoke allreduce. send " << x->numel() << " recv "
+            << out->numel();
+    PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
+        x->data<T>(), out->mutable_data<T>(ctx.GetPlace()), out->numel(),
+        NCCLTypeWrapper<T>::type, reduction_op_, comm->comms().at(idx),
+        ctx.cuda_device_context().stream()));
+    VLOG(3) << "gpu : "
+            << " finished allreduce. send " << x->numel() << " recv "
+            << out->numel();
   }
 };
 
@@ -94,13 +82,13 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                    "This kernel only runs on GPU device.");
-
-    auto ins = ctx.MultiInput<LoDTensor>("X");  // x0, x1, x2
-    auto outs = ctx.MultiOutput<LoDTensor>("Out");
-
+    auto x = ctx.Input<LoDTensor>("X");  // x0, x1, x2
+    auto out = ctx.Output<LoDTensor>("Out");
+    auto* comm = ctx.Input<Communicator>("Communicator");
+    int root = ctx.Attr<int>("root");
     std::string reduction = ctx.Attr<std::string>("reduction");
-    ncclRedOp_t reduction_op_ = ncclSum;
 
+    ncclRedOp_t reduction_op_ = ncclSum;
     if (reduction == "ncclMin") {
       reduction_op_ = ncclMin;
     } else if (reduction == "ncclMax") {
@@ -112,40 +100,21 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
     } else {
       PADDLE_THROW("Invalid reduction. default ncclSum.");
     }
-
-    int root = ctx.Attr<int>("root");
-    auto* comm = ctx.Input<Communicator>("Communicator");
-
-    auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
-                      ctx.device_context())
-                      .stream();
     // device id
     int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
-
-    auto ins_names = ctx.Inputs("X");
-    std::hash<std::string> hasher;
-    for (size_t i = 0; i < ins.size(); ++i) {
-      if (root == platform::kInvalidGPUId) {
-        root = hasher(ins_names[i]) % comm->comms().size();
-      }
-      T* recvbuffer = nullptr;
-      if (root == gpu_id) {
-        recvbuffer = outs[i]->mutable_data<T>(ctx.GetPlace());
-      }
-
-      VLOG(1) << "gpu : " << gpu_id << " invoke reduce. send "
-              << ins[i]->numel() << " recv " << outs[i]->numel();
-
-      PADDLE_ENFORCE(platform::dynload::ncclReduce(
-          ins[i]->data<T>(), recvbuffer, ins[i]->numel(),
-          NCCLTypeWrapper<T>::type, reduction_op_, root, comm->comms().at(idx),
-          stream));
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
-
-      VLOG(1) << "gpu : " << gpu_id << " finished reduce. send "
-              << ins[i]->numel() << " recv " << outs[i]->numel();
+    T* recvbuffer = nullptr;
+    if (root == gpu_id) {
+      recvbuffer = out->mutable_data<T>(ctx.GetPlace());
     }
+    VLOG(3) << "gpu : " << gpu_id << " invoke reduce. send " << x->numel()
+            << " recv " << out->numel();
+    PADDLE_ENFORCE(platform::dynload::ncclReduce(
+        x->data<T>(), recvbuffer, x->numel(), NCCLTypeWrapper<T>::type,
+        reduction_op_, root, comm->comms().at(idx),
+        ctx.cuda_device_context().stream()));
+    VLOG(3) << "gpu : " << gpu_id << " finished reduce. send " << x->numel()
+            << " recv " << out->numel();
   }
 };
 
@@ -155,47 +124,27 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                    "This kernel only runs on GPU device.");
-
     int root = ctx.Attr<int>("root");
-
     auto* comm = ctx.Input<Communicator>("Communicator");
-
-    auto stream = reinterpret_cast<const platform::CUDADeviceContext&>(
-                      ctx.device_context())
-                      .stream();
     // device id
     int gpu_id = boost::get<platform::CUDAPlace>(ctx.GetPlace()).GetDeviceId();
     int idx = comm->GetCommId(gpu_id);
-
     if (idx == root) {
-      auto ins = ctx.MultiInput<LoDTensor>("X");
-      for (size_t i = 0; i < ins.size(); ++i) {
-        VLOG(1) << "gpu : " << gpu_id << " invoke Bcast. send "
-                << ins[i]->numel();
-
-        VLOG(1) << " before ncclBcast";
-        PADDLE_ENFORCE(platform::dynload::ncclBcast(
-            (void*)ins[i]->data<T>(), ins[i]->numel(), NCCLTypeWrapper<T>::type,
-            root, comm->comms().at(idx), stream));
-        VLOG(1) << " after ncclBcast";
-        PADDLE_ENFORCE(cudaStreamSynchronize(stream));
-
-        VLOG(1) << "gpu : " << gpu_id << " finished Bcast.";
-      }
+      auto* x = ctx.Input<LoDTensor>("X");
+      VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. send " << x->numel();
+      PADDLE_ENFORCE(platform::dynload::ncclBcast(
+          (void*)x->data<T>(), x->numel(), NCCLTypeWrapper<T>::type, root,
+          comm->comms().at(idx), ctx.cuda_device_context().stream()));
+      VLOG(3) << "gpu : " << gpu_id << " finished Bcast.";
     } else {
-      auto outs = ctx.MultiOutput<LoDTensor>("Out");
-      for (size_t i = 0; i < outs.size(); ++i) {
-        VLOG(1) << "gpu : " << gpu_id << " invoke Bcast. recv buffer "
-                << framework::product(outs[i]->dims());
-
-        PADDLE_ENFORCE(platform::dynload::ncclBcast(
-            outs[i]->mutable_data<T>(ctx.GetPlace()), outs[i]->numel(),
-            NCCLTypeWrapper<T>::type, root, comm->comms().at(idx), stream));
-        PADDLE_ENFORCE(cudaStreamSynchronize(stream));
-
-        VLOG(1) << "gpu : " << gpu_id << " finished Bcast. recv "
-                << outs[i]->numel();
-      }
+      auto* out = ctx.Output<LoDTensor>("Out");
+      VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. recv buffer "
+              << framework::product(out->dims());
+      PADDLE_ENFORCE(platform::dynload::ncclBcast(
+          out->mutable_data<T>(ctx.GetPlace()), out->numel(),
+          NCCLTypeWrapper<T>::type, root, comm->comms().at(idx),
+          ctx.cuda_device_context().stream()));
+      VLOG(3) << "gpu : " << gpu_id << " finished Bcast. recv " << out->numel();
     }
   }
 };

paddle/fluid/operators/send_op.cc

@@ -88,6 +88,12 @@ class SendOp : public framework::OperatorBase {
         rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
       }
       PADDLE_ENFORCE(rpc_client->Wait());
+      // tell pservers that current trainer have called fetch
+      for (auto& ep : endpoints) {
+        VLOG(3) << "send fetch barrier, ep: " << ep;
+        rpc_client->AsyncSendFetchBarrier(ep);
+      }
+      PADDLE_ENFORCE(rpc_client->Wait());
     }
   }
 };

python/paddle/fluid/distribute_transpiler.py

@@ -250,6 +250,8 @@ class DistributeTranspiler:
     def get_trainer_program(self):
         # remove optimize ops and add a send op to main_program
         self.program.global_block().delete_ops(self.optimize_ops)
+        # FIXME(typhoonzero): serialize once will fix error occurs when clone.
+        self.program.__str__()
         return self.program
 
     def get_pserver_program(self, endpoint):
@@ -309,7 +311,8 @@ class DistributeTranspiler:
             for _, opt_op in enumerate(opt_op_on_pserver):
                 if ufind.is_connected(op, opt_op):
                     if self._is_opt_op(op):
-                        self._append_pserver_ops(optimize_block, op, endpoint)
+                        self._append_pserver_ops(optimize_block, op, endpoint,
+                                                 default_main_program())
                     else:
                         self._append_pserver_non_opt_ops(optimize_block, op)
                     break
@@ -520,7 +523,8 @@ class DistributeTranspiler:
             orig_var_name = varname[:suff_idx]
         return orig_var_name
 
-    def _append_pserver_ops(self, optimize_block, opt_op, endpoint):
+    def _append_pserver_ops(self, optimize_block, opt_op, endpoint,
+                            origin_program):
         program = optimize_block.program
         pserver_block = program.global_block()
         new_inputs = dict()
@@ -576,7 +580,17 @@ class DistributeTranspiler:
             elif key == "LearningRate":
                 # leraning rate variable has already be created by non-optimize op,
                 # don't create it once again.
-                new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]]
+                lr_varname = opt_op.input(key)[0]
+                if pserver_block.vars.has_key(lr_varname):
+                    new_inputs[key] = pserver_block.vars[opt_op.input(key)[0]]
+                else:
+                    origin_var = origin_program.global_block().vars[lr_varname]
+                    tmpvar = pserver_block.create_var(
+                        name=origin_var.name,
+                        persistable=origin_var.persistable,
+                        dtype=origin_var.dtype,
+                        shape=origin_var.shape)
+                    new_inputs[key] = tmpvar
 
         for key in opt_op.input_names:
             new_shape = None

python/paddle/fluid/tests/unittests/test_lookup_table_op.py

@@ -15,6 +15,8 @@
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
 
 
 class TestLookupTableOp(OpTest):
@@ -47,5 +49,52 @@ class TestLookupTableOpWithPadding(TestLookupTableOp):
         pass
 
 
+class TestLookupTableIdsIsSelectedRows(OpTest):
+    def check_with_place(self, place):
+        scope = core.Scope()
+
+        # create and initialize Variable
+        height = 10
+        rows = [0, 4, 4, 7]
+        row_numel = 12
+
+        # create and initialize W Variable
+        W = scope.var('W').get_tensor()
+        W_array = np.full((height, row_numel), 1.0).astype("float32")
+        for i in range(height):
+            W_array[i] *= i
+        W.set(W_array, place)
+
+        # create and initialize Ids Variable
+        ids_selected_rows = scope.var('Ids').get_selected_rows()
+        ids_selected_rows.set_height(len(rows))
+        ids_selected_rows.set_rows(rows)
+        np_array = np.ones((len(rows), row_numel)).astype("float32")
+        ids_tensor = ids_selected_rows.get_tensor()
+        ids_tensor.set(np_array, place)
+
+        # create Out Variable
+        Out = scope.var('Out').get_selected_rows()
+
+        # create and run lookup_table operator
+        concat_rows_op = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
+        concat_rows_op.run(scope, place)
+
+        # get result from Out
+        Out_tensor = Out.get_tensor()
+        result_array = np.array(Out_tensor)
+
+        # all(): return True if all elements of the iterable are true (or if the iterable is empty)
+        for idx, row in enumerate(rows):
+            assert (row == result_array[idx]).all()
+
+    def test_concat_rows(self):
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.check_with_place(place)
+
+
 if __name__ == "__main__":
     unittest.main()